Abstract:

Information entropy guarantees the most determinative probabilistic distribution of features in the original space, and it is able to deal with the problems such as value missing and noise; Manifold learning, i.e., Locally Linear Embedding in the dimensionalityreduced subspace, can completely present the relationship among features in the original manifoldstructured space. Combining both advantages, a new feature selection approach, named, Information Entropy based Locally Linear Embedding is proposed. Firstly, the feature information entropy is evaluated in the original space. Secondly, locally linear embedding is used to reduce the dimensionality in the feature subspace that keeps the most information. Finally, the feature subspace with lower dimensionality is obtained. In the given standard UCI dataset, the experimental results show the feasibility and validity of this method in feature selection.

Key words: information entropy;manifold learning;local linear embedding;dimensional reduction;classification